Methodology comparison for correcting woody component effects in leaf area index calculations from digital cover images in broadleaf forests

Abstract

Non-destructive methods are widely used for field measurement of leaf area index (LAI). However, the above-ground woody components of trees and shrubs, i.e., trunks and branches, largely affect the measured gap fraction, thus hindering the accurate measurement of LAI. Many efforts have been made to correct for the woody component effect and estimate LAI, but there is a lack of research to systematically compare and analyze current methods, which mainly include: 1) correcting using woody-to-total area ratio ($\alpha$), 2) transforming the leaf-on images into leaf-only images using artificially selected thresholds to determine whether woody components blocked leaves (TRM method), and 3) establishing the regression relationship between LAI and image features and/or other measured parameters using random forests (RFRM method) or 4) neural networks (NNRM method). We used rich data generated from 4734 scenes and 39 tree species to compare and analyze these methods. Additionally, considering the problems with the existing methods and the increasing requirements of LAI measurement, we proposed a new method (P2PLAI) using an image-to-image translation neural network (i.e., Pixel2Pixel) to transform the leaf-on images into leaf-only images. The effective LAI (LAIe) was estimated using the translated leaf-only images, and then the LAIe was converted into LAI using the clumping index. The results showed that the traditional method using $\alpha$ was limited by the accuracy of the $\alpha$ estimation, with the RMSE from 0.34 to 0.92 and the absolute percentage error (Bias%) from 9.56 % to 22.29 %. The TRM method could not stably and accurately transform the leaf-on images and underestimated LAI, with the RMSE from 0.13 to 0.78 and Bias% from 3.39 % to 21.13 %. The regression methods (i.e., RFRM and NNRM) had strong limitations since the accuracy of these two methods was related to the tree species and viewing zenith angles (VZAs) with RMSE up to 3.12 and Bias% up to 84.74 %. The P2PLAI method achieved the best agreement with the reference LAI. The RMSE and Bias% of P2PLAI respectively ranged from 0.05 to 0.26 and from 1.27 % to 7.70 % and were not influenced by tree species and VZAs. This study cautions against applying regression methods such as RFRM and NNRM for the indirect measurement of LAI in forests due to the complicated structures of vegetation components. The combination of an image-to-image translation neural network and a clumping effect correction model with physical meaning is recommended to measure LAI with digital photography.